I think the authors are using information theory to inappropriately flatten the complexity of the problem. On one hand we have “bits” of pre-processed sensory measurement data, then on the other hand we have “bits” of post-processed symbolic data: in many cases directly so via human language, but that would also include “the Terran unit moved a short distance” as a compact summary of a bunch of pixels updating in Sta…
Is it processing at a dozen bits per second, or hundreds of millions?
If the text the LLM generates is "that is true", can I consider that one bit of information?
I agree, they're artificially simplifying the framing of the question to generate a lower number than is sensible.